Managing an influence grid is like trying to unravel an enormous puzzle.
Grid operators must make sure that the correct amount of electricity flows to the correct areas at exactly the correct time, in a way that minimizes costs without overloading physical infrastructure. Furthermore, they need to unravel this complicated problem repeatedly and as quickly as possible to satisfy ever-changing demand.
To solve this ongoing puzzle, MIT researchers have developed a problem-solving tool that finds the optimal solution much faster than traditional approaches while ensuring that the answer doesn’t violate any of the system constraints. In an influence grid, restrictions might be, for instance, generator and line capability.
This recent tool integrates a feasibility search step with a robust machine learning model that’s trained to unravel the issue. The feasibility search step uses the model's prediction as a start line and iteratively refines the answer until one of the best possible answer is found.
The MIT system can solve complex problems persistently faster than traditional solvers while providing strong guarantees of success. For some extremely complex problems, it could find higher solutions than proven tools. The technology also outperformed purely machine learning approaches, that are fast but cannot at all times find workable solutions.
This recent tool not only helps plan electricity production on an influence grid, but is also applied to many forms of complicated problems, similar to developing recent products, managing investment portfolios, or planning production to satisfy consumer demand.
“To solve these particularly thorny problems well, we want to mix tools from machine learning, optimization, and electrical engineering to develop methods that make the correct trade-offs so as to add value to the domain while meeting its needs. You have to take a look at the needs of the appliance and design methods in a way that truly meets those needs,” says Priya Donti, Silverman Family Career Development Professor within the Department of Electrical Engineering and Computer Science (EECS) and principal investigator within the Laboratory for Information and Decision Systems (LID).
Donti, senior creator of an open access article Article about this recent tool called FSNetis accompanied by lead creator Hoang Nguyen, an EECS doctoral student. The paper might be presented on the Conference on Neural Information Processing Systems.
Combine approaches
Ensuring optimal power flow across an influence grid is a particularly difficult problem that’s becoming increasingly difficult for operators to unravel quickly.
“As we glance to integrate more renewable energy into the grid, operators must grapple with the incontrovertible fact that the quantity of power generation will fluctuate from moment to moment. At the identical time, many more distributed devices will have to be coordinated,” explains Donti.
Network operators often depend on traditional solvers that provide mathematical guarantees that the optimal solution doesn’t violate any problem constraints. However, these tools can take hours and even days to achieve an answer if the issue is especially complicated.
On the opposite hand, deep learning models can solve even very difficult problems in a fraction of the time, but the answer may ignore some essential limitations. For an influence grid operator, this may lead to problems similar to uncertain voltage levels and even grid failures.
“Machine learning models struggle to satisfy all constraints attributable to the numerous errors that occur throughout the training process,” explains Nguyen.
For FSNet, the researchers combined one of the best of each approaches in a two-stage problem-solving framework.
Focus on feasibility
In step one, a neural network predicts an answer to the optimization problem. Largely inspired by neurons within the human brain, neural networks are deep learning models that excel at recognizing patterns in data.
Next, a standard solver integrated with FSNet performs a proof-of-concept step. This optimization algorithm iteratively refines the initial prediction while ensuring that the answer doesn’t violate any constraints.
Because the feasibility search step relies on a mathematical model of the issue, it could possibly guarantee that the answer is possible.
“This step could be very essential. With FSNet we are able to get the strict guarantees that we want in practice,” says Hoang.
The researchers designed FSNet to take into consideration each essential forms of constraints (equality and inequality) at the identical time. This makes it easier to make use of than other approaches that will require customization of the neural network or a separate solution for every style of limitation.
“Here you’ll be able to just plug and play with different optimization solvers,” says Donti.
By pondering otherwise about how the neural network solves complex optimization problems, researchers were in a position to develop a brand new technique that works higher, she adds.
They compared FSNet to traditional solvers and pure machine learning approaches for a variety of difficult problems, including power grid optimization. Their system reduced solution times by orders of magnitude in comparison with baseline approaches while respecting all problem constraints.
FSNet has also found higher solutions to a few of the most difficult problems.
“This was surprising for us, but it surely makes perfect sense. Our neural network can discover an extra structure in the information that the unique optimization solver was not designed to use,” explains Donti.
In the longer term, researchers intend to make FSNet less memory-intensive, incorporate more efficient optimization algorithms, and scale it to deal with more realistic problems.
“Finding solutions to difficult optimization problems which might be feasible is paramount to finding those which might be near-optimal. Particularly for physical systems like power grids, near-optimal means nothing without feasibility. This work represents a very important step toward ensuring that deep learning models can produce predictions that satisfy constraints, with explicit guarantees for constraint enforcement,” says Kyri Baker, an associate professor on the University of Colorado Boulder, who was not involved on this work.
